We present a versatile unsupervised machine learning method for identifying minimum variance groups of molecules according to similarity in their dynamics or structures. We first consider molecular dynamics simulations of a collection of related systems, such as a set of proteins with different mutations. After constructing Markov state models for these simulations, we quantify the dynamical similarity of each pair of systems according to the Jensen-Shannon divergence between their Markovian transition probability matrices. By combining the divergence with hierarchical agglomerative clustering using Ward's minimum variance objective function, we can determine groups of systems based on their dynamics. This analysis enables an interpretable grouping of mutations with similar dynamical effects. Then, we apply the same method to common representations of small molecules for supervised machine learning tasks, such as SMILES strings and molecular fingerprints, in order to generate train/validation/test splits designed to avoid overfit models. By using the string edit distance and the Tanimoto coefficient to quantify the similarity between SMILES strings and molecular fingerprints, respectively, we identify minimum variance groups according to the chosen representation. We show that similarity according to molecular representation does not always align with heuristic similarity. Therefore, chemoinformatic supervised learning models may be at risk of overfitting if dataset partitioning for model cross-validation is performed according to heuristic class labels, since very similar representations may belong to the same class. This method provides insight into which properties are captured by the molecular representation, which enables an interpretable comparison of different representations.

Unsupervised clustering of molecules using minimum variance yields interpretable dynamical and structural groups

255th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.